Meta Research

Integrated gradients is a popular method for post-hoc model interpretability. Despite its popularity, the composition and relative impact of different regions of the integral are not well understood. We explore these effects and find that gradients in saturated regions of the scaling factor, where model output changes minimally, contribute disproportionately to the computed attribution. We propose a variant of Integrated Gradients which primarily captures gradients in unsaturated regions and evaluate this method on ImageNet classification networks. We find that this attribution technique shows higher model faithfulness and lower sensitivity to noise than standard Integrated Gradients.